Information Loss Problem Research Articles

A deep segmentation network for crack detection with progressive and hierarchical context fusion

Accurate detection of crack defects in infrastructure is crucial to ensure their safety and extend their service life. However, the presence of complex backgrounds, various shapes and sizes of crack defects, incomplete discontinuous crack defects, and class imbalance makes this task challenging. Traditional image processing techniques are sensitive to image noise and may miss cracks due to weak textures, complex lighting conditions, and other similar objects on the pavement. In recent years, deep learning-based segmentation networks have been proposed for crack defect detection, but they still face challenges in high-precision crack segmentation due to insufficient local feature processing, information loss caused by pooling operations, and limited receptive fields. To address these issues, we propose an end-to-end deep crack segmentation network, called PHCF-Net, which incorporates progressive and hierarchical context fusion. Firstly, the proposed network consists of progressive context fusion (PCF) and hierarchical context fusion (HCF) blocks for effective aggregation of global and local context information. Secondly, a multi-scale context fusion (MCF) block is proposed for multi-scale context extraction and aggregation. Finally, in order to solve the information loss problem caused by pooling operations, a hierarchical context fusion (HCF) block is proposed for effective aggregation of the deep and shallow features. In addition, a multi-scale input unit is also applied to the proposed segmentation network to obtain more context information. To evaluate the performance of PHCF-Net, we have conducted experiments on two publicly available crack segmentation datasets and compared its performance with mainstream segmentation models. The results demonstrate that proposed PHCF-Net achieves better pixel-level crack detection results and outperforms other advanced segmentation models.

Information loss, mixing and emergent type III1 factors

A manifestation of the black hole information loss problem is that the two-point function of probe operators in a large Anti-de Sitter black hole decays in time, whereas, on the boundary CFT, it is expected to be an almost periodic function of time. We point out that the decay of the two-point function (clustering in time) holds important clues to the nature of observable algebras, states, and dynamics in quantum gravity.We call operators that cluster in time “mixing” and explore the necessary and sufficient conditions for mixing. The information loss problem is a special case of the statement that in type I algebras, there exists no mixing operators. We prove that, in a thermofield double state (KMS state), if mixing operators form an algebra (close under multiplication), the resulting algebra must be a von Neumann type III1 factor. In other words, the physically intuitive requirement that all nonconserved operators should exponentially mix is so strong that it fixes the observable algebra to be an exotic algebra called a type III1 factor. More generally, for an arbitrary out-of-equilibrium state of a general quantum system (von Neumann algebra), we show that if the set of operators that mix under modular flow forms an algebra, it is a type III1 von Neumann factor.In a theory of Generalized Free Fields (GFF), we show that if the two-point function clusters in time, all operators are mixing, and the algebra is a type III1 factor. For example, in \U0001d4a9 = 4 SYM, above the Hawking-Page phase transition, clustering of the single trace operators implies that the algebra is a type III1 factor, settling a recent conjecture of Leutheusser and Liu. We explicitly construct the C∗-algebra and von Neumann subalgebras of GFF associated with time bands and, more generally, open sets of the bulk spacetime using the HKLL reconstruction map.

HimGNN: a novel hierarchical molecular graph representation learning framework for property prediction.

Accurate prediction of molecular properties is an important topic in drug discovery. Recent works have developed various representation schemes for molecular structures to capture different chemical information in molecules. The atom and motif can be viewed as hierarchical molecular structures that are widely used for learning molecular representations to predict chemical properties. Previous works have attempted to exploit both atom and motif to address the problem of information loss in single representation learning for various tasks. To further fuse such hierarchical information, the correspondence between learned chemical features from different molecular structures should be considered. Herein, we propose a novel framework for molecular property prediction, called hierarchical molecular graph neural networks (HimGNN). HimGNN learns hierarchical topology representations by applying graph neural networks on atom- and motif-based graphs. In order to boost the representational power of the motif feature, we design a Transformer-based local augmentation module to enrich motif features by introducing heterogeneous atom information in motif representation learning. Besides, we focus on the molecular hierarchical relationship and propose a simple yet effective rescaling module, called contextual self-rescaling, that adaptively recalibrates molecular representations by explicitly modelling interdependencies between atom and motif features. Extensive computational experiments demonstrate that HimGNN can achieve promising performances over state-of-the-art baselines on both classification and regression tasks in molecular property prediction.

Peripheral Blood Leukocyte Detection Based on an Improved Detection Transformer Algorithm.

The combination of a blood cell analyzer and artificial microscopy to detect white blood cells is used in hospitals. Blood cell analyzers not only have large throughput, but they also cannot detect cell morphology; although artificial microscopy has high accuracy, it is inefficient and prone to missed detections. In view of the above problems, a method based on Fourier ptychographic microscopy (FPM) and deep learning to detect peripheral blood leukocytes is proposed in this paper. Firstly, high-resolution and wide-field microscopic images of human peripheral blood cells are obtained using the FPM system, and the cell image data are enhanced with DCGANs (deep convolution generative adversarial networks) to construct datasets for performance evaluation. Then, an improved DETR (detection transformer) algorithm is proposed to improve the detection accuracy of small white blood cell targets; that is, the residual module Conv Block in the feature extraction part of the DETR network is improved to reduce the problem of information loss caused by downsampling. Finally, CIOU (complete intersection over union) is introduced as the bounding box loss function, which avoids the problem that GIOU (generalized intersection over union) is difficult to optimize when the two boxes are far away and the convergence speed is faster. The experimental results show that the mAP of the improved DETR algorithm in the detection of human peripheral white blood cells is 0.936. In addition, this algorithm is compared with other convolutional neural networks in terms of average accuracy, parameters, and number of inference frames per second, which verifies the feasibility of this method in microscopic medical image detection.

An Online Quality Detection Method With Ensemble Learning on Imbalance Data for Wave Soldering

Abstract Online detection of wave soldering is an important method of inspecting defective products in the workshop. Accurate quality detection can reduce production costs and provide support for quality warnings in the wave soldering process. However, there are still problems in improving the detection accuracy for the defect class. Although class imbalance in data can be addressed by data-level methods such as over-sampling and under-sampling, these methods destroy the integrity of the original data set and may cause information loss and over-fitting problems. In order to solve the above problems, this article focuses on how to design a new loss function that fuses class weights from focal loss (FS) and sample weights from AdaBoost to improve attention to the minority samples without changing data distribution. In this way, an FS-AdaBoost-RegNet model based on transfer learning is constructed to enhance the detection accuracy in industrial environment. Finally, the images of the wave soldering from an electronic assembly workshop are taken to validate the performance of the proposed method. The experiment on 941 testing samples of the imbalance datasets showed that the FS-AdaBoost-RegNet model with new loss function reached the overall accuracy of 98.39%, and the overall recall of 96.19%. The results proved that the proposed method promotes the ability to identify defect class compared with other methods.

A lightweight multi-sensory field-based dual-feature fusion residual network for bird song recognition

Bird song recognition plays an important function in ecosystem balance monitoring, biodiversity detection, and biodiversity conservation. Due to the complexity of the natural environment, based on deep learning, there is a problem of information loss in extracting audio features with a single filter. Identifying bird sounds efficiently and quickly is still a challenge. To address this problem, a lightweight multi-sensory field dual-feature fusion residual network (LDFSRE-NET) is proposed in this paper. Firstly, using the feature extraction filters based on Mel and SincNet to extract birdsong’s low-frequency and timbre information. The proposed dual-feature fusion module (FFMS) is used to fuse the low-frequency and timbre information with the differences between the two feature sets. Secondly, the double-layer residual module (DBNet), connected by basicblock and downblock, is used as the backbone network for bird song recognition to improve the training speed. To improve the different perceptual fields of the backbone network, the 3 × 3 convolutional modules in the basicblocks of the two residual modules are replaced with a Diverse Branch Block. They make the network performs better on recognition tasks under complex situations of multiple branches. Then, the ShuffleAttention attention module is embedded between the two layers of the residual module for transferring its valid information, enhancing the spectrogram ripple feature, and further improving the network’s recognition performance. Finally, extensive experiments are conducted on three datasets: the self-built 30-class bird song dataset (Birdselfdata), the public datasets Birdsdata and Urbansound8K. The model proposed in this paper surpasses the state-of-the-art sound recognition model methods in terms of efficiency and accuracy. The recognition accuracy on these three datasets of this model are 96.75%, 96.46%, and 97.98%, with the F1-score of 96.79%, 96.39%, and 97.88%.

An end-to-end multiple side-outputs fusion deep supervision network based remote sensing image change detection algorithm

Change detection is an effective means of monitoring surface changes. In order to solve the problem of reducing detection accuracy in remote sensing image change detection methods that only focus on extracting deep semantic information while ignoring high-resolution shallow information, this paper proposes an end-to-end multi-side output fusion deep supervised network for high-resolution biphasic remote sensing image change detection. The enhanced feature extraction module is used to increase the Receptive field and capture multi-scale features. The dense skip connection module utilizes skip connections to fuse shallow position information and deep semantic information, alleviating the problem of feature information loss. The deep supervision module with multiple side-outputs fusion utilizes the information complementarity between different scale feature maps, combined with multi-scale prediction maps, to improve the robustness of target scale changes. The combined Loss function is used to alleviate the problem of imbalance between positive and negative samples, so that more attention is paid to the learning of changing characteristics in the process of training the network. The proposed method was validated on CDD dataset and DSIFN dataset, with F1 scores of 91.58% and 86.69%, respectively, which were improved by 1.00% and 5.53% compared to the suboptimal network.

An improved YOLO v4 used for grape detection in unstructured environment.

Visual recognition is the most critical function of a harvesting robot, and the accuracy of the harvesting action is based on the performance of visual recognition. However, unstructured environment, such as severe occlusion, fruits overlap, illumination changes, complex backgrounds, and even heavy fog weather, pose series of serious challenges to the detection accuracy of the recognition algorithm. Hence, this paper proposes an improved YOLO v4 model, called YOLO v4+, to cope with the challenges brought by unstructured environment. The output of each Resblock_body in the backbone is processed using a simple, parameterless attention mechanism for full dimensional refinement of extracted features. Further, in order to alleviate the problem of feature information loss, a multi scale feature fusion module with fusion weight and jump connection structure was pro-posed. In addition, the focal loss function is adopted and the hyperparameters α, γ are adjusted to 0.75 and 2. The experimental results show that the average precision of the YOLO v4+ model is 94.25% and the F1 score is 93%, which is 3.35% and 3% higher than the original YOLO v4 respectively. Compared with several state-of-the-art detection models, YOLO v4+ not only has the highest comprehensive ability, but also has better generalization ability. Selecting the corresponding augmentation method for specific working condition can greatly improve the model detection accuracy. Applying the proposed method to harvesting robots may enhance the applicability and robustness of the robotic system.

CSCIM_FS: Cosine similarity coefficient and information measurement criterion-based feature selection method for high-dimensional data

Feature selection (FS) based on mutual information (MI) metrics needs to discretize the data in preprocessing, which is a convenient way to identify correlation between features. However, information loss often occurs in data discretization. In order to solve this information loss problem, this paper proposes a FS algorithm based on cosine similarity coefficient and information measurement criterion (CSCIM_FS). First, the MI between features and tags is calculated, and features are sorted out according to the MI calculated. Then, a feature matrix is constructed to transform the one-dimensional feature sequence into a two-dimensional square matrix. Next, cosine transform is adopted to obtain the high-frequency components of the feature matrix, and sampling is conducted to derive the hash fingerprint of the feature matrix. After that, the similarity between every two features is calculated on the basis of the hash fingerprints of different features. Finally, the feature weight is calculated according to tags, the MI and similarity between features, and a key feature subset is obtained and used to conduct feature selection from the data. The experimental results on several UCI public datasets show that CSCIM_FS algorithm selected a feature subset with high accuracy, and that this algorithm performs better than MIM, CMIM, mRMR and other algorithms.

An efficient 3D object detection method based on Fast Guided Anchor Stereo RCNN

In most binocular 3D detection algorithms, a large number of anchor points need to be selected, which leads to the problem of slow feature extraction. To solve this problem, an anchor-guided 3D object detection algorithm for autonomous driving is proposed based on Stereo Recurrent Convolutional Neutral Network (Stereo RCNN), which is called Fast Guided Anchored Stereo RCNN (FGAS RCNN). The proposed FGAS framework is divided into two stages. In the first stage, a probability map is generated for the left and right input images to determine the foreground position. Sparse anchor points and corresponding sparse anchor boxes are generated from the prior information. Left and right anchors are used as a whole to generate a 2D preselection box. In the second stage, a Feature Pyramid Network (FPN) based on key point generation network is used to generate key points, which are combined with stereo regression to generate 3D preselected boxes. Finally, instance-level disparity estimation is proposed to solve the problem of pixel-level information loss in the original image. Instance-level disparity is combined with instance segmentation masks to improve the accuracy of center depth on the 3D bounding box. Extensive experiments on the challenging Kitti dataset and NuScences dataset show that the proposed method reduces the computational cost while maintaining a high regression rate without any depth information and prior information of position. Compared to other methods, the proposed method has higher efficiency, better robustness and stronger generalization ability.

DnRCNN: Deep Recurrent Convolutional Neural Network for HSI Destriping.

In spite of achieving promising results in hyperspectral image (HSI) restoration, deep-learning-based methodologies still face the problem of spectral or spatial information loss due to neglecting the inner correlation of HSI. To address this issue, we propose an innovative deep recurrent convolution neural network (DnRCNN) model for HSI destriping. To the best of our knowledge, this is the first study on HSI destriping from the perspective of inner band and interband correlation explorations with the recurrent convolution neural network. In the novel DnRCNN, a selective recurrent memory unit (SRMU) is designed to respectively extract the correlative features involved in spectral and spatial domains. Moreover, an innovative recurrent fusion (RF) strategy incorporated with group concatenation is further proposed to remove strip noise and preserve scene details using the complementary features from SRMU. Experimental results on extensive HSI datasets validated that the proposed method achieves a new state-of-the-art (SOTA) HSI destriping performance.

Underwater Target Detection Algorithm Based on Feature Fusion Enhancement

Underwater robots that use optical images for dynamic target detection often encounter image blurring, poor contrast, and indistinct target features. As a result, the underwater robots have poor detection performance with a high rate of missed detections. To overcome these issues, a feature-enhanced algorithm for underwater target detection has been proposed in this paper. Based on YOLOv7, a feature enhancement module utilizing a triple-attention mechanism is developed to improve the network’s feature extraction ability without increasing the computational or algorithmic parameter quantity. Moreover, comprehensively considering the impact of a redundant feature in the images on detection accuracy, the ASPPCSPC structure was built. A parallel spatial convolutional pooling structure based on the original feature pyramid fusion structure, SPPCSPC, is introduced. The GhostNet network was utilized to optimize its convolution module, which reduces the model’s parameter quantity and optimizes the feature map. Furthermore, a Cat-BiFPN structure was designed to address the problem of fine-grained information loss in YOLOv7 feature fusion by adopting a weighted nonlinear fusion strategy to enhance the algorithm’s adaptability. Using the UPRC offshore dataset for validation, the algorithm’s detection accuracy was increased by 2.9%, and the recall rate was improved by 2.3% compared to the original YOLOv7 algorithm. In addition, the model quantity is reduced by 11.2%, and the model size is compressed by 10.9%. The experimental results significantly establish the validity of the proposed algorithm.

BiDFDC-Net: a dense connection network based on bi-directional feedback for skin image segmentation.

Accurate segmentation of skin lesions in dermoscopic images plays an important role in improving the survival rate of patients. However, due to the blurred boundaries of pigment regions, the diversity of lesion features, and the mutations and metastases of diseased cells, the effectiveness and robustness of skin image segmentation algorithms are still a challenging subject. For this reason, we proposed a bi-directional feedback dense connection network framework (called BiDFDC-Net), which can perform skin lesions accurately. Firstly, under the framework of U-Net, we integrated the edge modules into each layer of the encoder which can solve the problem of gradient vanishing and network information loss caused by network deepening. Then, each layer of our model takes input from the previous layer and passes its feature map to the densely connected network of subsequent layers to achieve information interaction and enhance feature propagation and reuse. Finally, in the decoder stage, a two-branch module was used to feed the dense feedback branch and the ordinary feedback branch back to the same layer of coding, to realize the fusion of multi-scale features and multi-level context information. By testing on the two datasets of ISIC-2018 and PH2, the accuracy on the two datasets was given by 93.51% and 94.58%, respectively.

NSNet: An N-Shaped Convolutional Neural Network with Multi-Scale Information for Image Denoising

Deep learning models with convolutional operators have received widespread attention for their good image denoising performance. However, since the convolutional operation prefers to extract local features, the extracted features may lose some global information, such as texture, structure, and color characteristics, when the object in the image is large. To address this issue, this paper proposes an N-shaped convolutional neural network with the ability to extract multi-scale features to capture more useful information and alleviate the problem of global information loss. The proposed network has two main parts: a multi-scale input layer and a multi-scale feature extraction layer. The former uses a two-dimensional Haar wavelet to create an image pyramid, which contains the corrupted image’s high- and low-frequency components at different scales. The latter uses a U-shaped convolutional network to extract features at different scales from this image pyramid. The method sets the mean-squared error as the loss function and uses the residual learning strategy to learn the image noise directly. Compared with some existing image denoising methods, the proposed method shows good performance in gray and color image denoising, especially in textures and contours.

Multi-scale attention and dilation network for small defect detection

Although object detection methods have got surprising performance in simple natural scenes, it is still a challenging task to apply them to complex scenes, especially when the detected objects are small which is common in defect detection tasks of industrial scenarios. Motivated by the higher resolution feature maps, the better detection performance on small objects, we increase the resolution of the feature layers in this paper, and introduce convolutional block attention module (CBAM) into the feature aggregation network to better integrate features at different scales and make better use of the information of small defects. For solving the problem of global context information loss caused by limiting the perceptual field by increasing the resolution of the feature layers, we use dilated convolution blocks to expand the perceptual field. Experimental results show that our method improves the mAP(0.5) on the Industrial Surface Defect Dataset from 88.3% in Yolov5 to 89.2%, mAP(0.75) on the DAGM 2007 increases from 76.56% in Yolov5 to 77.65%, mAP on small object increases by 2.6% on Industrial Surface Defect Dataset and 5.8% on DAGM 2007, which proves the effectiveness of our method for small defect detection.

Lightweight Spatial-Spectral Network Based on 3D-2D Multi-Group Feature Extraction Module for Hyperspectral Image Classification

ABSTRACT Each pixel in the hyperspectral image contains tens to hundreds of narrow bands of detailed spectral information captured by a hyperspectral sensor. Hyperspectral image classification (HSIC) is widely used in remote sensing image analysis. Convolutional neural network (CNN) is one of the most commonly used deep learning methods based on visual data processing. In recent years, hyperspectral classification method based on the convolutional neural network has shown excellent performance. However, single-scale feature extraction may lose some important detail information and cannot guarantee that the best spatial features are captured. Multi-scale character extraction and increasingly deep network structures can improve classification accuracy, but also bring a large number of computational parameters, resulting in higher computational costs. In response to this series of problems, this paper proposes a lightweight spatial-spectral network based on 3D − 2D multi-group feature extraction module (MGFM) for HSCI. In the 3D-MGFM, the input feature maps are grouped and processed in parallel, and then the information of each channel is integrated by point-wise convolution to achieve spatial-spectral feature fusion. We use dilated convolutions with different dilation rate on the convolution kernel to improve the problem of feature information loss, so that we can better learn deep-level spatial-spectral features. In the 2D-MGFM, we selectively emphasize spectral feature information through SENet and extract multi-scale features based on Depthwise Separable Convolution (DSC) grouping. Experimental results on four hyperspectral datasets demonstrate that the proposed method achieves better classification performance than other state-of-the-art methods.

Boundary-constrained interpretable image reconstruction network for deep compressive sensing

Compressive Sensing (CS) networks have received increasing attention due to their wide applications in data compression, single pixel imaging, medical imaging, etc. However, there are several difficult problems for CS reconstruction. First, many early CS methods treat deep networks as black boxes. Second, some of the newest methods combine iterative optimizers with neural networks to make their architecture explainable, but their performances need to be further improved. Third, most CS reconstruction methods based on deep learning ignore the significance of edge information, greatly limiting the representation ability of their reconstruction models. To this end, this paper proposes a boundary-constrained CS reconstruction model, and we cast a boundary-constrained CS reconstruction problem as two subproblems, which can be alternatively optimized. The alternatively-iterative optimization process can be expanded into an Edge Guided Interpretable image compressive sensing Network (EGINet) for image reconstruction. The proposed EGINet has three modules: an edge-aware feature-extraction module, an edge-guided intermediate-variable updating module and an intermediate-variable guided image reconstruction module. To resolve the problem of information loss in the iterative process, we introduce a multiple-memory enhancement mechanism to explore feature dependency across different iteration phases of EGINet. Meanwhile, we design a parallel-cross fusion module to selectively fuse boundary features and image features. A large number of experimental results have shown that, compared with both black-box and explainable CS reconstruction methods, the proposed EGINet has better reconstruction performances in terms of SSIM and PSNR, especially on recovering image boundary information, while maintaining network interpretability.

ST-YOLOA: a Swin-transformer-based YOLO model with an attention mechanism for SAR ship detection under complex background.

A synthetic aperture radar (SAR) image is crucial for ship detection in computer vision. Due to the background clutter, pose variations, and scale changes, it is a challenge to construct a SAR ship detection model with low false-alarm rates and high accuracy. Therefore, this paper proposes a novel SAR ship detection model called ST-YOLOA. First, the Swin Transformer network architecture and coordinate attention (CA) model are embedded in the STCNet backbone network to enhance the feature extraction performance and capture global information. Second, we used the PANet path aggregation network with a residual structure to construct the feature pyramid to increase global feature extraction capability. Next, to cope with the local interference and semantic information loss problems, a novel up/down-sampling method is proposed. Finally, the decoupled detection head is used to achieve the predicted output of the target position and the boundary box to improve convergence speed and detection accuracy. To demonstrate the efficiency of the proposed method, we have constructed three SAR ship detection datasets: a norm test set (NTS), a complex test set (CTS), and a merged test set (MTS). The experimental results show that our ST-YOLOA achieved an accuracy of 97.37%, 75.69%, and 88.50% on the three datasets, respectively, superior to the effects of other state-of-the-art methods. Our ST-YOLOA performs favorably in complex scenarios, and the accuracy is 4.83% higher than YOLOX on the CTS. Moreover, ST-YOLOA achieves real-time detection with a speed of 21.4 FPS.

SADENet: A supervised attention delicate enhanced network for subtle person detection

Object detection algorithms have been developed for decades, and one of remaining important challenges is the detection of subtle targets in complex environments. Many convolution neural network (CNN) based methods have been proposed to handle this problem. However, the information pertaining to subtle targets is delicate and can easily be lost or neglected in the upsampling and downsampling stages of conventional CNN. In this paper, we propose a new subtle target detection network, named the supervised attention delicate enhanced network (SADENet). Firstly, two sample vertebras are designed to solve the problem of information loss during upsampling and downsampling. Then, an enhanced context attention module (ECAM) is proposed, which extends the original supervised heatmap by fusing information from different scales. Based on the ECAM and the respective advantages of high-resolution and low-resolution feature maps, two effective sampling attention modules, the semantic fidelity upsampling module and the location enhancement downsampling module, are created to build a bi-directional connection neck structure. The effectiveness of our proposed method is demonstrated in experiments on the TinyPerson dataset, which outperforms many state-of-the-art detectors, especially in the detection of tiny and tiny1 scale targets.

Reconstructing Ocean Subsurface Temperature and Salinity from Sea Surface Information Based on Dual Path Convolutional Neural Networks

Satellite remote sensing can provide observation information of the sea surface, and using the sea surface information to reconstruct the subsurface temperature (ST) and subsurface salinity (SS) information has significant application values. This study proposes an intelligent algorithm based on Dual Path Convolutional Neural Networks (DP-CNNs) to reconstruct the ST and SS. The DP-CNN can integrate known information including sea surface temperature (SST), sea surface salinity (SSS), and sea surface height (SSH) to reconstruct the ST and SS. The reconstruction model based on DP-CNN can solve the problem of detail information loss in traditional CNN (Convolutional Neural Network) models. This study performs experiments for the South China Sea under different seasons using reanalysis data. The experimental results show that the DP-CNN models have higher reconstruction accuracy than the CNN models, and this proves that DP-CNNs effectively mitigate the loss of detailed information in the CNN models. Compared with the ground truth data, the ST/SS reconstruction results of the DP-CNN model exhibited a high coefficient of determination (0.93/0.86) and a low root mean square error (around 0.31 °C/0.05 PSU). Therefore, the DP-CNN models can be used as an effective approach to reconstruct ST and SS using sea surface information.